Former Google DeepMind researcher Danijar Hafner is building a stealth AI startup focused on agents that can plan ahead using predictive world models, according to a new MIT Technology Review report. Separate reporting identifies the company as Embo, co-founded with fellow former DeepMind researcher Wilson Yan, with robotics emerging as its first major application.
The interesting part is not another AI agent promising to reason through a longer list of steps. Hafner’s previous work points towards a different architecture: teach a model how an environment changes, let an agent simulate possible actions inside that learned world, then use those imagined outcomes to choose what to do. If Embo can make that approach reliable outside research benchmarks, it could reduce one of robotics’ highest costs: learning through repeated physical trial and error.
| What we know | Current position |
|---|---|
| Startup | Embo, a stealth-stage AI company |
| Founders | Danijar Hafner and Wilson Yan, both former Google DeepMind researchers |
| Technical focus | World models and agents that can predict consequences before acting |
| Initial application | Robotics and embodied AI |
| Funding | Earlier reporting said Embo was in talks to raise more than $100 million in seed funding. A completed round has not been publicly confirmed. |
| Public product | No generally available API, SDK, pricing or customer product has been announced in the material reviewed |
The new report fills in a missing piece around Embo
Hafner’s departure from Google DeepMind was already known, as was his connection to a new stealth company. Earlier reporting by The Information identified that company as Embo and said Hafner and Yan were developing world models specifically for robotics. It also reported that Embo was in talks for a seed round of more than $100 million led by Andreessen Horowitz, with Khosla Ventures, DST Global and Striker Venture Partners also discussing participation.
The financing needs careful wording. It was reported as being in talks, not as a completed funding announcement. Embo remains unusually quiet for a company linked to a potential nine-figure seed round, with little public information about its product, customers, deployment model, or commercial timetable.
The latest report is more revealing technically because it frames the work around agents capable of planning ahead. That fits Hafner’s research history far better than a generic description such as “AI for robotics”. His work has repeatedly investigated how agents can learn an internal model of an environment and use it to reason about future consequences.
Planning ahead here is not just longer AI reasoning
Most current AI agents plan at the instruction level. A model might decide to search for information, open a tool, inspect the result and choose the next step. That is useful, but it does not necessarily mean the system has learned how the environment itself behaves.
A world-model agent tries to learn that missing layer. Given a current state and a possible action, the model predicts what could happen next. The agent can then evaluate several imagined futures before committing to an action in the real environment.
Hafner’s Dreamer 4 most clearly shows where this approach can lead. Dreamer 4 learns a predictive model of Minecraft, trains behaviour inside that model through reinforcement learning and then executes the learned policy in the actual game. The researchers report that it can obtain diamonds from offline data alone, a task requiring sequences of more than 20,000 mouse and keyboard actions.
The research also reported substantially lower data use than OpenAI’s VPT agent in that specific Minecraft comparison. That result should not be turned into a general claim that world models make robotics 100 times more data-efficient, but it demonstrates why the approach attracts attention: expensive environment interaction can potentially be replaced with much larger amounts of simulated experience.
Embo’s opportunity is really about the economics of robot learning
Language models benefited from a vast amount of already-digitised training material. Robots do not have the same advantage. A useful robot trajectory can require hardware, cameras, operators, a controlled environment, resets after failures and careful recording of the actions that produced each state.
Physical mistakes also have costs that a chatbot does not face. A failed robot action can drop an object, damage equipment or require a person to reset the workspace. Collecting another million words of text and collecting another million high-quality robot interactions are completely different engineering problems.
A sufficiently accurate world model changes that equation. Real interaction is still needed to teach and validate the model, but policies can practise additional behaviours inside the learned environment. One physical trajectory can therefore contribute to more than one learning pass.
A second, less-discussed advantage is counterfactual data. A robot normally observes only the result of the action it actually performed. A world model can potentially evaluate what might have happened under several alternative actions from the same starting state. For planning, that ability may be more valuable than simply generating another realistic-looking video.
The hard problem is action fidelity, not making convincing video
This is also where world-model demonstrations can be misleading. A generated environment can look extremely convincing while still being a poor place to train an agent.
A recurring lesson from independent Dreamer-style reproductions is that visual quality and controllability can separate. A model may generate plausible motion yet react weakly or incorrectly to the action it was given. Longer simulated sequences introduce another failure mode: small prediction errors compound until the imagined state gradually stops matching a physically plausible future.
For an ordinary video generator, a minor error several seconds into a clip might be tolerable. For an agent, the same error can corrupt the decision that follows it. If the simulated cupboard is open when the real cupboard would still be closed, every subsequent action may be planned against the wrong world.
This means we should not judge Embo primarily on photorealistic demonstrations. The stronger evidence would show that actions reliably cause the correct state changes and that useful planning survives increasingly long horizons.
Five benchmarks would tell us whether Embo has something commercially important
The usual image or video generation metrics would reveal relatively little about an agent intended to make decisions. We would want to test Embo as a planning system.
| Test | What it should measure | Failure to watch for |
|---|---|---|
| Action fidelity | Whether a requested action produces the correct change in the simulated environment | Visually plausible output that largely ignores the action |
| Long-horizon stability | How long the world remains coherent across repeated predicted actions | Small state errors accumulating until planning becomes unreliable |
| Counterfactual accuracy | Whether different candidate actions create meaningfully different and correct futures | The model collapsing towards similar likely-looking outcomes regardless of the action |
| World-to-real transfer | Whether policies trained in imagination still work on physical hardware | An agent exploiting inaccuracies that exist only inside the learned model |
| Recovery from model error | Whether the agent can incorporate a new observation when reality differs from its prediction | Continuing a plan built around a state that no longer exists |
The last test could prove particularly important. Perfect world models are unrealistic. A commercially useful agent needs to recognise when its prediction was wrong, observe the actual state again and repair its plan rather than blindly completing an obsolete sequence.
This could eventually change software agents as well as robots
Embo is most directly associated with robotics, and there is no evidence yet of a general computer-use product. The underlying planning problem is broader, though.
A computer agent also operates inside a changing environment. Clicking a button changes application state. Editing a file changes what later commands will see. Authentication can expire. A failed command can create a different starting point for every step that follows. Our explanation of generative AI vs agentic AI covers the practical shift that occurs once a model moves from producing an answer to acting inside external systems.
A useful learned model of software state could eventually let an agent simulate possible actions rather than discovering every failure by taking the action for real. The environment differs greatly from robotics, but the evaluation principle is similar: a planner is valuable only if its imagined consequences are accurate enough to improve completed work.
This is why our AI coding agent benchmark focuses on accepted outcomes, retries and execution rather than treating underlying model intelligence as a substitute for successful agent behaviour. A world-model agent would need the same discipline. Better predictions are useful only if they increase successful task completion.
Several important details about Embo are still missing
The research pedigree is unusually strong, but Embo is still closer to a technical thesis than a product a developer can evaluate. Several questions need answering before its commercial significance is clear:
- Will Embo sell foundation models, an agent runtime, a robotics development platform or complete trained policies?
- Which robot embodiments and sensor configurations will it support first?
- How much action-labelled data does a new hardware platform require before the world model becomes useful?
- Can one world model transfer knowledge between different robots, or will substantial per-platform training still be necessary?
- How quickly can the system correct itself after encountering an environment outside its training distribution?
- What does inference cost when an agent evaluates several possible futures before taking every important action?
- Has the previously reported funding round closed, and on what terms?
The business model depends heavily on the answers. A general model that can be adapted to many robot platforms is infrastructure. A system that needs extensive new training for each robot, factory and task behaves more like a specialised robotics project, with very different margins and deployment times.
DIY AI view: Embo needs to prove planning when its imagined world is wrong
Hafner and Yan have already helped demonstrate that world models can support surprisingly long sequences of behaviour in controlled environments. The startup’s harder challenge is turning that research direction into something developers can trust when reality refuses to follow the prediction.
The most impressive future Embo demo would therefore not be a flawless robot completing a rehearsed task. It would be a robot that begins with a plan, encounters something the model predicted incorrectly, recognises the mismatch, and recovers without a human rebuilding the workflow.
If Embo can make that loop reliable, planning-capable agents become more than agents with longer reasoning traces. They become systems that can test possible futures, choose between them, observe what actually happened, and revise the plan. That would meaningfully change how autonomous AI is trained and evaluated.
Until Embo publishes a product, technical documentation or deployment results, the commercial claims should remain provisional. The reason to watch it is much narrower and stronger: two researchers closely associated with the Dreamer line appear to be turning imagination-based agent training into a company, and robotics gives them one of the clearest places to prove whether the idea works outside a research paper.
